Distance Measures for Image Segmentation Evaluation
- PDF / 930,315 Bytes
- 10 Pages / 600.03 x 792 pts Page_size
- 71 Downloads / 243 Views
Distance Measures for Image Segmentation Evaluation Xiaoyi Jiang,1 Cyril Marti,2 Christophe Irniger,2 and Horst Bunke2 1 Computer
Vision and Pattern Recognition Group, Department of Computer Science, University of M¨unster, Einsteinstrasse 62, D-48149 M¨unster, Germany 2 Institute of Computer Science and Applied Mathematics, University of Bern, Neubr¨ uckstrasse 10, CH-3012 Bern, Switzerland Received 17 March 2005; Revised 10 July 2005; Accepted 31 July 2005 The task considered in this paper is performance evaluation of region segmentation algorithms in the ground-truth-based paradigm. Given a machine segmentation and a ground-truth segmentation, performance measures are needed. We propose to consider the image segmentation problem as one of data clustering and, as a consequence, to use measures for comparing clusterings developed in statistics and machine learning. By doing so, we obtain a variety of performance measures which have not been used before in image processing. In particular, some of these measures have the highly desired property of being a metric. Experimental results are reported on both synthetic and real data to validate the measures and compare them with others. Copyright © 2006 Hindawi Publishing Corporation. All rights reserved.
1.
INTRODUCTION
Image segmentation and recognition are central problems of image processing for which we do not yet have any general purpose solution approaching human-level competence. Recognition is basically a classification task and one can empirically estimate the recognition performance (probability of misclassification) by counting classification errors on a test set. Today, reporting recognition performance on large data sets is a well-accepted standard. In contrast, segmentation performance evaluation remains subjective. Typically, results on a few images are shown and the authors argue why they look good. The readers frequently do not know whether the results have been opportunistically selected or are typical examples, and how well the demonstrated performance extrapolates to larger sets of images. The main challenge is that the question “to what extent is this segmentation correct” is much more subtle than “is this face from person x.” While a huge number of segmentation algorithms have been reported, there is only little work on methodologies of segmentation performance evaluation [1]. Several segmentation tasks can be identified: edge detection, region segmentation, and detection of curvilinear structures. Their performance evaluation is of quite different nature. For instance, an evaluation of detection algorithms for curvilinear structures must take the elongated shape of this particular feature into account [2]. In some sense, edge detection and region segmentation are two dual problems and their performance evaluation appears to be a similar task. One may convert a segmented region map to an equivalent edge map by
marking the region boundaries only and then applying any edge detection evaluation method. However, a simple example, as shown in Figure 1
Data Loading...
